基于PCNN和矩特征的遥感图像目标识别研究

发布时间：2018-11-16 17:35

【摘要】：近年来,基于卫星遥感图像的目标识别技术得到迅猛发展,被广泛地应用于军事领域和民用领域。目前,各国学者在遥感图像飞机识别上取得了一定的突破。然而,现实环境远不及理论来的理想化,遥感图像必然存在噪声、复杂背景等干扰,这定当会影响后续的识别,因此现有的理论成果仍有不足之处,比如在识别精度、耗时量、通用性等方面还不尽如人意。为此,如何在复杂的环境中高效的识别出飞机目标成为了本文的研究重点和关键。众所周知,目标识别过程大致包括:预处理、分割、特征提取及识别。而本文的研究重点——遥感图像分割、特征提取,已取得如下成果:1、为了提高遥感图像目标分割的精度,提出了一种基于引力搜索算法参数优化的改进脉冲耦合神经网络(PCNN-Pulse Coupled Neural Network)遥感图像分割算法。首先,通过二次描述神经元间的激励和抑制关系,改进连接输入项和动态阈值来优化经典PCNN模型。然后利用上述模型对输入信息进行点火处理,并从其输出结果中提取图像熵和能量的比值作为引力搜索算法的适应度函数,且将熵的变化值作为引力搜索算法的收敛依据,利用引力搜索算法的全局搜索能力寻找PCNN模型中影响分割效果的关键参数的最优值。最终将该算法与OTSU、最大熵直方图算法和原始PCNN算法进行对比,并通过Matlab仿真实验证明了本文算法更适用于遥感图像分割。2、针对几何不变矩对仿射形变目标描述的不足,为提高飞机类型的识别精度,给出了基于小波和仿射不变矩特征融合的飞机识别算法。首先对二值飞机图像做归一化操作,并分别计算归一化飞机目标的小波矩和仿射不变矩特征值;然后通过计算样本特征均值与标准差的商,筛选出鲁棒性好、稳定性高的特征,通过归一化方法进行融合;最后将五种不同型号的飞机构造成样本集,并采用支持向量机(Support Vector Machine,SVM)方法识别测试样本的型号。实验将不同类型的矩特征、不同容量的样本集就识别精度、稳定性指标进行了比较,结果表明,文中给出的方法提高了精度,而且在训练样本集较小时仍能获得较高的识别率。3、基于以上两个重要步骤,再结合支持向量机,完整地完成了整个识别过程。通过实验证明文中提出的方法不仅能克服类型各异及比例不同的噪声干扰,还能适用于复杂背景图像的飞机目标。同时,也保证了较高的识别精度和较少的耗时量。
[Abstract]:In recent years, the technology of target recognition based on satellite remote sensing image has been developed rapidly and widely used in military and civilian fields. At present, many scholars have made a breakthrough in remote sensing image aircraft recognition. However, the real environment is far from idealized by theory, and there must be noise and complex background interference in remote sensing images, which will definitely affect the subsequent recognition. Therefore, the existing theoretical achievements still have some shortcomings, such as recognition accuracy, time consuming, etc. Versatility and other aspects are not satisfactory. Therefore, how to efficiently identify aircraft targets in complex environments has become the focus and key of this paper. As we all know, the process of target recognition includes preprocessing, segmentation, feature extraction and recognition. The research focus of this paper, remote sensing image segmentation, feature extraction, has achieved the following results: 1, in order to improve the accuracy of remote sensing image segmentation, An improved pulse coupled neural network (PCNN-Pulse Coupled Neural Network) algorithm for remote sensing image segmentation based on parameter optimization of gravity search algorithm is proposed. Firstly, the classical PCNN model is optimized by quadratic description of the excitation and suppression relationship between neurons and the improvement of connecting input terms and dynamic threshold. Then the input information is ignited by the above model, and the ratio of image entropy and energy is extracted from the output result as the fitness function of the gravity search algorithm, and the change of entropy is taken as the convergence basis of the gravity search algorithm. The global search ability of the gravitational search algorithm is used to find the optimal value of the key parameters in the PCNN model that affect the segmentation effect. Finally, the algorithm is compared with the OTSU, maximum entropy histogram algorithm and the original PCNN algorithm, and the Matlab simulation results show that the proposed algorithm is more suitable for remote sensing image segmentation. 2. In order to improve the accuracy of aircraft type recognition, an aircraft recognition algorithm based on wavelet and affine moment invariant feature fusion is presented. Firstly, the binary plane image is normalized, and the eigenvalues of wavelet moment and affine invariant moment of the normalized aircraft target are calculated respectively. Then, by calculating the quotient of the mean and standard deviation of the sample feature, the features with good robustness and high stability are screened out, and the fusion is carried out by the normalization method. Finally, five different types of flying mechanism are made into sample sets, and support vector machine (Support Vector Machine,SVM) method is used to identify the model of test samples. The experimental results show that the method proposed in this paper has improved the accuracy and stability of the samples with different types of moment features and different capacity. Moreover, a high recognition rate can be obtained when the training sample set is small. 3. Based on the above two important steps and combining with support vector machine, the whole recognition process is completed. It is proved by experiments that the proposed method can not only overcome the noise interference of different types and proportions, but also be suitable for aircraft targets with complex background images. At the same time, it also ensures higher recognition accuracy and less time consuming.
【学位授予单位】：江苏科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】